Integrating Deep Learning and Radiomics for Osteoporosis Screening in Elderly Population using Lumbar CT

Abstract

Rationale and Objectives:To create and validate an all-encompassing method that combines deep learning and radiomics, enabling the utilization of routine lumbar CT scans for opportunistic screening of osteoporosis. Materials and Methods:This research collected patient data retrospectively from January 2020 through December 2022. A sample of 100 lumbar vertebrae was selected to develop a UPerNet model for segmenting vertebral bone marrow. The remaining vertebrae were utilized as validation data for the segmentation model and employed to generate a radiomic signature for osteoporosis diagnosis. Subsequently, the remaining vertebrae were allocated into a training set, an internal validation set, and an external validation set, following a 3:1:1 ratio. A total of 1794 radiomic features were extracted from the lumbar vertebral bone marrow. Feature selection was sequentially carried out using the minimum-redundancy maximum-relevance (mRMR) and then the least absolute shrinkage and selection operator (LASSO), followed by the construction of the radiomic signature using logistic regression. The performance of the vertebral segmentation model was evaluated with the Dice coefficient. Intraclass correlation coefficients (ICCs) were calculated to assess the consistency of radiomic feature extraction from automatic segmentation by the UPerNet model and manual segmentation by radiologists. The diagnostic performance of the radiomic signature was assessed using receiver operating characteristic (ROC) analysis. Results: This study encompassed 438 lumbar vertebrae from 127 patients, with 168 of these vertebrae being osteoporotic. The UPerNet model achieved a Dice coefficient of 0.90 (95%CI: 0.84-0.95) for validation. Of the 1794 radiomic features extracted, 88.45% showed ICC values over 0.8. The area under the curve (AUC) for radiomic signature in the external validation set reached 0.96 (95%CI: 0.91-1.00). Conclusions: The radiomic signature derived from automatically segmented CT images of lumbar bone marrow using the UPerNet model exhibited high accuracy in osteoporosis screening.